Backport PR #52470 on branch 2.0.x (BUG: describe not returning ArrowDtype) (#52495)

Backport PR #52470: BUG: describe not returning ArrowDtype

Author: mroeschke

doc/source/whatsnew/v2.0.1.rst

@@ -21,6 +21,7 @@ Fixed regressions
 Bug fixes
 ~~~~~~~~~
 - Fixed bug in :func:`merge` when merging with ``ArrowDtype`` one one and a NumPy dtype on the other side (:issue:`52406`)
+- Bug in :meth:`Series.describe` not returning :class:`ArrowDtype` with ``pyarrow.float64`` type with numeric data (:issue:`52427`)
 
 .. ---------------------------------------------------------------------------
 .. _whatsnew_201.other:

pandas/core/methods/describe.py

@@ -37,7 +37,8 @@
     is_timedelta64_dtype,
 )
 
-import pandas as pd
+from pandas.core.arrays.arrow.dtype import ArrowDtype
+from pandas.core.arrays.floating import Float64Dtype
 from pandas.core.reshape.concat import concat
 
 from pandas.io.formats.format import format_percentiles
@@ -230,7 +231,12 @@ def describe_numeric_1d(series: Series, percentiles: Sequence[float]) -> Series:
     # GH#48340 - always return float on non-complex numeric data
     dtype: DtypeObj | None
     if is_extension_array_dtype(series):
-        dtype = pd.Float64Dtype()
+        if isinstance(series.dtype, ArrowDtype):
+            import pyarrow as pa
+
+            dtype = ArrowDtype(pa.float64())
+        else:
+            dtype = Float64Dtype()
     elif is_numeric_dtype(series) and not is_complex_dtype(series):
         dtype = np.dtype("float")
     else:

pandas/tests/extension/test_arrow.py

@@ -2343,3 +2343,16 @@ def test_setitem_boolean_replace_with_mask_segfault():
     expected = arr.copy()
     arr[np.zeros((N,), dtype=np.bool_)] = False
     assert arr._data == expected._data
+
+
+@pytest.mark.parametrize("pa_type", tm.ALL_INT_PYARROW_DTYPES + tm.FLOAT_PYARROW_DTYPES)
+def test_describe_numeric_data(pa_type):
+    # GH 52470
+    data = pd.Series([1, 2, 3], dtype=ArrowDtype(pa_type))
+    result = data.describe()
+    expected = pd.Series(
+        [3, 2, 1, 1, 1.5, 2.0, 2.5, 3],
+        dtype=ArrowDtype(pa.float64()),
+        index=["count", "mean", "std", "min", "25%", "50%", "75%", "max"],
+    )
+    tm.assert_series_equal(result, expected)